A Data Processing Guide in the Big Data Jungle

Too many choices? Don’t get lost!

We are deep in the Big Data jungle. According to Gartner’s Hype Cycle for Emerging Technologies, Big Data has now officially passed the “peak of inflated expectations”, and is now on a one-way trip to the “trough of disillusionment”. Gartner says it’s done so rather fast, because we already have consistency in the way we approach this technology, and because most new advances are additive rather than revolutionary.

Pig, Hive, Impala, Tez and Spark: which one suits for which use case?

With so much hype and so many new advances, it’s easy to get lost. This little guide gives you an overview on data processing technologies in the Big Data jungle and tries to identify the best use cases for each.

  • Pig: Pig is often useful for pulling apart unstructured and nested data like text or JSON. Since Pig Latin is a procedural language, it is a very good choice for developing data pipelines on Hadoop. Pig is based on MapReduce and has tools for data storage, data execution and data manipulation.
  • Hive: Hive was original “relational on Hadoop” and is the first Hadoop SQL (HiveQL to be precise) query engine. Hive is still the most mature engine from all in this guide, as well as the slowest one. Hive is also based on MapReduce and is a very good choice for heavy ETL tasks where reliability is important, eg. daily aggregation jobs.
  • Impala: Impala is the only native open-source SQL query engine in the Hadoop world. It skips MapReduce entirely and is best used for SQL queries over big volumes. Impala is also capable of delivering results interactively over bigger volumes and with a much faster speed than other Hadoop query engines.
  • Tez: Tez may be considered as a better and faster base for query engines like Pig and Hive. Tez gets around limitations imposed by MapReduce and enables use cases with near-real-time performance and Machine Learning, which do not fit well into the MapReduce paradigm.
  • Spark: Spark is an in-memory query engine that also skips MapReduce. Perfect use cases for Spark are streaming, interactive data processing and ad-hoc analysis of moderate-sized data sets (as big as the cluster’s RAM). The ability of Spark to reuse data in-memory is the real highlight for these use cases. Spark SQL offers relational connectivity.